In [1]:
from keras.models import load_model
from keras.datasets import mnist
from keras import backend as K
from keras.utils.np_utils import to_categorical
import matplotlib.pyplot as plt
import numpy as np
%matplotlib inline
In [2]:
var_path = 'resultados/network_1/'
Load no dataset original, apenas para pegar um exemplo de imagem sem nenhum pre-processamento.
In [3]:
dataset = mnist.load_data()
Essas informações abaixo devem ser consistentes com o modelo treinado.
In [4]:
(X_treino, y_treino), (X_teste, y_teste) = mnist.load_data()
qtd_linhas = 28
qtd_colunas = 28
numero_classes = 10 # temos de 0 a 9 no dataset
num_filtros = 32
X_treino = X_treino.reshape(X_treino.shape[0], qtd_linhas, qtd_colunas, 1).astype('float')
X_teste = X_teste.reshape(X_teste.shape[0], qtd_linhas, qtd_colunas, 1).astype('float')
y_treino = to_categorical(y_treino, numero_classes)
y_teste = to_categorical(y_teste, numero_classes)
Ler o modelo treinado¶
In [5]:
modelo = load_model(var_path + '/model_trained.h5')
modelo.load_weights(var_path + '/weights.h5')
modelo.evaluate(X_teste, y_teste)
Out[5]:
Definição da visualização do modelo¶
In [6]:
from keras.utils.visualize_util import plot
for layer in modelo.layers:
print layer.name
plot(modelo,to_file=var_path + '/model.png',show_shapes=True,show_layer_names=True)
plt.xticks(np.array([]))
plt.yticks(np.array([]))
plt.imshow(plt.imread(var_path + '/model.png'))
Out[6]:
Visualizando filtros¶
In [7]:
img = np.zeros((28,28))
for a in xrange(28):
for b in xrange(28):
img[a][b] = X_treino[0][a,b]
for i in xrange(len(modelo.layers)): # os filtros estao onde estao os layers convolucionais
m0 = modelo.layers[i]
name = m0.name
if "convolution" not in name:
continue
print name
ws = m0.get_weights()
#print np.shape(ws[0])
flt = np.zeros((np.shape(ws[0])[0],np.shape(ws[0])[1]))
fig=plt.figure(figsize=(8,8))
for oz in xrange(np.shape(ws[0])[2]):
for z in xrange(np.shape(ws[0])[3]):
ax = fig.add_subplot(6, 6, z+1)
sa = np.shape(ws[0])[0]
sb = np.shape(ws[0])[1]
for a in xrange(sa):
for b in xrange(sb):
flt[a][b] = ws[0][a,b,oz,z]
plt.xticks(np.array([]))
plt.yticks(np.array([]))
plt.tight_layout()
ax.imshow(flt,interpolation='nearest', cmap=plt.cm.gray)
plt.show()
In [8]:
for num in xrange(10):
for num_camada_para_ser_visualizada in xrange(len(modelo.layers)):
print "="*20 + "camada " + str(num_camada_para_ser_visualizada) + "="*20
print modelo.layers[num_camada_para_ser_visualizada].name
output_layer = modelo.layers[num_camada_para_ser_visualizada].output
output_fn = K.function([modelo.layers[0].input], [output_layer])
output_image = output_fn([X_treino[num:num+1]])[0]
#print "shape: " + str(np.shape(output_image))
fig=plt.figure(figsize=(8,8))
for i in range(num_filtros):
if len(np.shape(output_image))==4:
ax = fig.add_subplot(6, 6, i+1)
ax.imshow(output_image[0,:,:,i],interpolation='nearest', cmap=plt.cm.gray) #to see the first filter
plt.xticks(np.array([]))
plt.yticks(np.array([]))
plt.tight_layout()
elif len(np.shape(output_image))==2:
ax = fig.add_subplot(1,1,i+1)
s=1*np.shape(output_image)[1]/8
output_image = np.transpose(np.repeat(output_image[0],s).reshape((len(output_image[0]),s)))
ax.imshow(output_image[:],interpolation='nearest', cmap=plt.cm.gray) #to see the first filter
plt.xticks(np.array([]))
plt.yticks(np.array([]))
plt.tight_layout()
break
plt.show()
In [ ]: